Automatic Video Analysis for Fisheries Survey Systems

Thesis (Ph.D.)--University of Washington, 2015

Fisheries survey with the use of cameras has drawn increasing attention since it enables a non-lethal and high-resolution sampling of the fish stocks. This dissertation strives to build an automatic video analysis system that provides an effective solution to video-based fisheries surveys. The proposed system includes segmentation and tracking with controlled background, segmentation and tracking with uncontrolled background, body length estimation, and species recognition. In segmentation and tracking with controlled background, a reliable algorithm for low-contrast and low-frame-rate (LFR) stereo videos is proposed for trawl-based camera systems. The automatic segmentation integrates the histogram backprojection to double local thresholding to produce accurate shape boundary. Built upon a feature-based temporal matching method, the multiple-target tracking algorithm via Viterbi data association is proposed to overcome the abrupt motion and frequent entrance/exit of targets. In segmentation and tracking with uncontrolled background, a novel tracking algorithm based on deformable multiple kernels (DMK) is proposed. In light of the deformable part model (DPM), a number of kernels are defined to represent the holistic body and parts of the object. Kernel motion is efficiently estimated by the mean-shift algorithm on color and texture features. The DPM deformation costs are adopted as soft constraints on kernel positions to maintain the part configuration when matching HOG features. After tracking, an active contour algorithm is applied around each target to realize target segmentation. In body length estimation, an efficient block-matching approach performs successful stereo matching. It also enables an automatic fish-body tail compensation to reduce segmentation error and allows for an accurate fish length measurement after disparity refinement and stereo triangulation. In species recognition, an innovative object recognition framework is proposed. To extract discriminative fish species features, both the supervised feature descriptors and the unsupervised feature learning method via the novel non-rigid part model are presented. For the classifier, unsupervised clustering generates a binary class hierarchy, where each node is a classifier. Partial classification is introduced to assign coarse labels to ambiguous data by systematically optimizing the “benefit” of classifier indecision. Several experiments show the effectiveness of the proposed system against public datasets and practical scenarios.